diff options
Diffstat (limited to 'pkg/sentry/mm')
-rw-r--r-- | pkg/sentry/mm/BUILD | 3 | ||||
-rw-r--r-- | pkg/sentry/mm/address_space.go | 52 | ||||
-rw-r--r-- | pkg/sentry/mm/aio_context.go | 101 | ||||
-rw-r--r-- | pkg/sentry/mm/aio_context_state.go | 2 | ||||
-rw-r--r-- | pkg/sentry/mm/lifecycle.go | 38 | ||||
-rw-r--r-- | pkg/sentry/mm/metadata.go | 24 | ||||
-rw-r--r-- | pkg/sentry/mm/mm.go | 15 | ||||
-rw-r--r-- | pkg/sentry/mm/mm_test.go | 2 | ||||
-rw-r--r-- | pkg/sentry/mm/procfs.go | 4 | ||||
-rw-r--r-- | pkg/sentry/mm/syscalls.go | 4 | ||||
-rw-r--r-- | pkg/sentry/mm/vma.go | 4 |
11 files changed, 172 insertions, 77 deletions
diff --git a/pkg/sentry/mm/BUILD b/pkg/sentry/mm/BUILD index e5729ced5..a036ce53c 100644 --- a/pkg/sentry/mm/BUILD +++ b/pkg/sentry/mm/BUILD @@ -25,6 +25,7 @@ go_template_instance( out = "vma_set.go", consts = { "minDegree": "8", + "trackGaps": "1", }, imports = { "usermem": "gvisor.dev/gvisor/pkg/usermem", @@ -105,8 +106,8 @@ go_library( "//pkg/safecopy", "//pkg/safemem", "//pkg/sentry/arch", - "//pkg/sentry/fs", "//pkg/sentry/fs/proc/seqfile", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/futex", "//pkg/sentry/kernel/shm", diff --git a/pkg/sentry/mm/address_space.go b/pkg/sentry/mm/address_space.go index e58a63deb..5c667117c 100644 --- a/pkg/sentry/mm/address_space.go +++ b/pkg/sentry/mm/address_space.go @@ -18,7 +18,7 @@ import ( "fmt" "sync/atomic" - "gvisor.dev/gvisor/pkg/atomicbitops" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/usermem" ) @@ -39,11 +39,18 @@ func (mm *MemoryManager) AddressSpace() platform.AddressSpace { // // When this MemoryManager is no longer needed by a task, it should call // Deactivate to release the reference. -func (mm *MemoryManager) Activate() error { +func (mm *MemoryManager) Activate(ctx context.Context) error { // Fast path: the MemoryManager already has an active // platform.AddressSpace, and we just need to indicate that we need it too. - if atomicbitops.IncUnlessZeroInt32(&mm.active) { - return nil + for { + active := atomic.LoadInt32(&mm.active) + if active == 0 { + // Fall back to the slow path. + break + } + if atomic.CompareAndSwapInt32(&mm.active, active, active+1) { + return nil + } } for { @@ -85,16 +92,20 @@ func (mm *MemoryManager) Activate() error { if as == nil { // AddressSpace is unavailable, we must wait. // - // activeMu must not be held while waiting, as the user - // of the address space we are waiting on may attempt - // to take activeMu. - // - // Don't call UninterruptibleSleepStart to register the - // wait to allow the watchdog stuck task to trigger in - // case a process is starved waiting for the address - // space. + // activeMu must not be held while waiting, as the user of the address + // space we are waiting on may attempt to take activeMu. mm.activeMu.Unlock() + + sleep := mm.p.CooperativelySchedulesAddressSpace() && mm.sleepForActivation + if sleep { + // Mark this task sleeping while waiting for the address space to + // prevent the watchdog from reporting it as a stuck task. + ctx.UninterruptibleSleepStart(false) + } <-c + if sleep { + ctx.UninterruptibleSleepFinish(false) + } continue } @@ -118,8 +129,15 @@ func (mm *MemoryManager) Activate() error { func (mm *MemoryManager) Deactivate() { // Fast path: this is not the last goroutine to deactivate the // MemoryManager. - if atomicbitops.DecUnlessOneInt32(&mm.active) { - return + for { + active := atomic.LoadInt32(&mm.active) + if active == 1 { + // Fall back to the slow path. + break + } + if atomic.CompareAndSwapInt32(&mm.active, active, active-1) { + return + } } mm.activeMu.Lock() @@ -183,8 +201,10 @@ func (mm *MemoryManager) mapASLocked(pseg pmaIterator, ar usermem.AddrRange, pre if pma.needCOW { perms.Write = false } - if err := mm.as.MapFile(pmaMapAR.Start, pma.file, pseg.fileRangeOf(pmaMapAR), perms, precommit); err != nil { - return err + if perms.Any() { // MapFile precondition + if err := mm.as.MapFile(pmaMapAR.Start, pma.file, pseg.fileRangeOf(pmaMapAR), perms, precommit); err != nil { + return err + } } pseg = pseg.NextSegment() } diff --git a/pkg/sentry/mm/aio_context.go b/pkg/sentry/mm/aio_context.go index cb29d94b0..379148903 100644 --- a/pkg/sentry/mm/aio_context.go +++ b/pkg/sentry/mm/aio_context.go @@ -59,25 +59,27 @@ func (a *aioManager) newAIOContext(events uint32, id uint64) bool { } a.contexts[id] = &AIOContext{ - done: make(chan struct{}, 1), + requestReady: make(chan struct{}, 1), maxOutstanding: events, } return true } -// destroyAIOContext destroys an asynchronous I/O context. +// destroyAIOContext destroys an asynchronous I/O context. It doesn't wait for +// for pending requests to complete. Returns the destroyed AIOContext so it can +// be drained. // -// False is returned if the context does not exist. -func (a *aioManager) destroyAIOContext(id uint64) bool { +// Nil is returned if the context does not exist. +func (a *aioManager) destroyAIOContext(id uint64) *AIOContext { a.mu.Lock() defer a.mu.Unlock() ctx, ok := a.contexts[id] if !ok { - return false + return nil } delete(a.contexts, id) ctx.destroy() - return true + return ctx } // lookupAIOContext looks up the given context. @@ -102,8 +104,8 @@ type ioResult struct { // // +stateify savable type AIOContext struct { - // done is the notification channel used for all requests. - done chan struct{} `state:"nosave"` + // requestReady is the notification channel used for all requests. + requestReady chan struct{} `state:"nosave"` // mu protects below. mu sync.Mutex `state:"nosave"` @@ -129,8 +131,14 @@ func (ctx *AIOContext) destroy() { ctx.mu.Lock() defer ctx.mu.Unlock() ctx.dead = true - if ctx.outstanding == 0 { - close(ctx.done) + ctx.checkForDone() +} + +// Preconditions: ctx.mu must be held by caller. +func (ctx *AIOContext) checkForDone() { + if ctx.dead && ctx.outstanding == 0 { + close(ctx.requestReady) + ctx.requestReady = nil } } @@ -154,11 +162,12 @@ func (ctx *AIOContext) PopRequest() (interface{}, bool) { // Is there anything ready? if e := ctx.results.Front(); e != nil { - ctx.results.Remove(e) - ctx.outstanding-- - if ctx.outstanding == 0 && ctx.dead { - close(ctx.done) + if ctx.outstanding == 0 { + panic("AIOContext outstanding is going negative") } + ctx.outstanding-- + ctx.results.Remove(e) + ctx.checkForDone() return e.data, true } return nil, false @@ -172,26 +181,58 @@ func (ctx *AIOContext) FinishRequest(data interface{}) { // Push to the list and notify opportunistically. The channel notify // here is guaranteed to be safe because outstanding must be non-zero. - // The done channel is only closed when outstanding reaches zero. + // The requestReady channel is only closed when outstanding reaches zero. ctx.results.PushBack(&ioResult{data: data}) select { - case ctx.done <- struct{}{}: + case ctx.requestReady <- struct{}{}: default: } } // WaitChannel returns a channel that is notified when an AIO request is -// completed. -// -// The boolean return value indicates whether or not the context is active. -func (ctx *AIOContext) WaitChannel() (chan struct{}, bool) { +// completed. Returns nil if the context is destroyed and there are no more +// outstanding requests. +func (ctx *AIOContext) WaitChannel() chan struct{} { ctx.mu.Lock() defer ctx.mu.Unlock() - if ctx.outstanding == 0 && ctx.dead { - return nil, false + return ctx.requestReady +} + +// Dead returns true if the context has been destroyed. +func (ctx *AIOContext) Dead() bool { + ctx.mu.Lock() + defer ctx.mu.Unlock() + return ctx.dead +} + +// CancelPendingRequest forgets about a request that hasn't yet completed. +func (ctx *AIOContext) CancelPendingRequest() { + ctx.mu.Lock() + defer ctx.mu.Unlock() + + if ctx.outstanding == 0 { + panic("AIOContext outstanding is going negative") } - return ctx.done, true + ctx.outstanding-- + ctx.checkForDone() +} + +// Drain drops all completed requests. Pending requests remain untouched. +func (ctx *AIOContext) Drain() { + ctx.mu.Lock() + defer ctx.mu.Unlock() + + if ctx.outstanding == 0 { + return + } + size := uint32(ctx.results.Len()) + if ctx.outstanding < size { + panic("AIOContext outstanding is going negative") + } + ctx.outstanding -= size + ctx.results.Reset() + ctx.checkForDone() } // aioMappable implements memmap.MappingIdentity and memmap.Mappable for AIO @@ -332,9 +373,9 @@ func (mm *MemoryManager) NewAIOContext(ctx context.Context, events uint32) (uint Length: aioRingBufferSize, MappingIdentity: m, Mappable: m, - // TODO(fvoznika): Linux does "do_mmap_pgoff(..., PROT_READ | - // PROT_WRITE, ...)" in fs/aio.c:aio_setup_ring(); why do we make this - // mapping read-only? + // Linux uses "do_mmap_pgoff(..., PROT_READ | PROT_WRITE, ...)" in + // fs/aio.c:aio_setup_ring(). Since we don't implement AIO_RING_MAGIC, + // user mode should not write to this page. Perms: usermem.Read, MaxPerms: usermem.Read, }) @@ -349,11 +390,11 @@ func (mm *MemoryManager) NewAIOContext(ctx context.Context, events uint32) (uint return id, nil } -// DestroyAIOContext destroys an asynchronous I/O context. It returns false if -// the context does not exist. -func (mm *MemoryManager) DestroyAIOContext(ctx context.Context, id uint64) bool { +// DestroyAIOContext destroys an asynchronous I/O context. It returns the +// destroyed context. nil if the context does not exist. +func (mm *MemoryManager) DestroyAIOContext(ctx context.Context, id uint64) *AIOContext { if _, ok := mm.LookupAIOContext(ctx, id); !ok { - return false + return nil } // Only unmaps after it assured that the address is a valid aio context to diff --git a/pkg/sentry/mm/aio_context_state.go b/pkg/sentry/mm/aio_context_state.go index c37fc9f7b..3dabac1af 100644 --- a/pkg/sentry/mm/aio_context_state.go +++ b/pkg/sentry/mm/aio_context_state.go @@ -16,5 +16,5 @@ package mm // afterLoad is invoked by stateify. func (a *AIOContext) afterLoad() { - a.done = make(chan struct{}, 1) + a.requestReady = make(chan struct{}, 1) } diff --git a/pkg/sentry/mm/lifecycle.go b/pkg/sentry/mm/lifecycle.go index 47b8fbf43..aac56679b 100644 --- a/pkg/sentry/mm/lifecycle.go +++ b/pkg/sentry/mm/lifecycle.go @@ -18,7 +18,6 @@ import ( "fmt" "sync/atomic" - "gvisor.dev/gvisor/pkg/atomicbitops" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/limits" @@ -29,16 +28,17 @@ import ( ) // NewMemoryManager returns a new MemoryManager with no mappings and 1 user. -func NewMemoryManager(p platform.Platform, mfp pgalloc.MemoryFileProvider) *MemoryManager { +func NewMemoryManager(p platform.Platform, mfp pgalloc.MemoryFileProvider, sleepForActivation bool) *MemoryManager { return &MemoryManager{ - p: p, - mfp: mfp, - haveASIO: p.SupportsAddressSpaceIO(), - privateRefs: &privateRefs{}, - users: 1, - auxv: arch.Auxv{}, - dumpability: UserDumpable, - aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + p: p, + mfp: mfp, + haveASIO: p.SupportsAddressSpaceIO(), + privateRefs: &privateRefs{}, + users: 1, + auxv: arch.Auxv{}, + dumpability: UserDumpable, + aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + sleepForActivation: sleepForActivation, } } @@ -80,9 +80,11 @@ func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) { envv: mm.envv, auxv: append(arch.Auxv(nil), mm.auxv...), // IncRef'd below, once we know that there isn't an error. - executable: mm.executable, - dumpability: mm.dumpability, - aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + executable: mm.executable, + dumpability: mm.dumpability, + aioManager: aioManager{contexts: make(map[uint64]*AIOContext)}, + sleepForActivation: mm.sleepForActivation, + vdsoSigReturnAddr: mm.vdsoSigReturnAddr, } // Copy vmas. @@ -229,7 +231,15 @@ func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) { // IncUsers increments mm's user count and returns true. If the user count is // already 0, IncUsers does nothing and returns false. func (mm *MemoryManager) IncUsers() bool { - return atomicbitops.IncUnlessZeroInt32(&mm.users) + for { + users := atomic.LoadInt32(&mm.users) + if users == 0 { + return false + } + if atomic.CompareAndSwapInt32(&mm.users, users, users+1) { + return true + } + } } // DecUsers decrements mm's user count. If the user count reaches 0, all diff --git a/pkg/sentry/mm/metadata.go b/pkg/sentry/mm/metadata.go index f550acae0..28e5057f7 100644 --- a/pkg/sentry/mm/metadata.go +++ b/pkg/sentry/mm/metadata.go @@ -16,7 +16,7 @@ package mm import ( "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/usermem" ) @@ -132,7 +132,7 @@ func (mm *MemoryManager) SetAuxv(auxv arch.Auxv) { // // An additional reference will be taken in the case of a non-nil executable, // which must be released by the caller. -func (mm *MemoryManager) Executable() *fs.Dirent { +func (mm *MemoryManager) Executable() fsbridge.File { mm.metadataMu.Lock() defer mm.metadataMu.Unlock() @@ -147,15 +147,15 @@ func (mm *MemoryManager) Executable() *fs.Dirent { // SetExecutable sets the executable. // // This takes a reference on d. -func (mm *MemoryManager) SetExecutable(d *fs.Dirent) { +func (mm *MemoryManager) SetExecutable(file fsbridge.File) { mm.metadataMu.Lock() // Grab a new reference. - d.IncRef() + file.IncRef() // Set the executable. orig := mm.executable - mm.executable = d + mm.executable = file mm.metadataMu.Unlock() @@ -167,3 +167,17 @@ func (mm *MemoryManager) SetExecutable(d *fs.Dirent) { orig.DecRef() } } + +// VDSOSigReturn returns the address of vdso_sigreturn. +func (mm *MemoryManager) VDSOSigReturn() uint64 { + mm.metadataMu.Lock() + defer mm.metadataMu.Unlock() + return mm.vdsoSigReturnAddr +} + +// SetVDSOSigReturn sets the address of vdso_sigreturn. +func (mm *MemoryManager) SetVDSOSigReturn(addr uint64) { + mm.metadataMu.Lock() + defer mm.metadataMu.Unlock() + mm.vdsoSigReturnAddr = addr +} diff --git a/pkg/sentry/mm/mm.go b/pkg/sentry/mm/mm.go index 09e582dd3..6db7c3d40 100644 --- a/pkg/sentry/mm/mm.go +++ b/pkg/sentry/mm/mm.go @@ -35,9 +35,10 @@ package mm import ( + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/platform" @@ -215,7 +216,7 @@ type MemoryManager struct { // is not nil, it holds a reference on the Dirent. // // executable is protected by metadataMu. - executable *fs.Dirent + executable fsbridge.File // dumpability describes if and how this MemoryManager may be dumped to // userspace. @@ -226,6 +227,14 @@ type MemoryManager struct { // aioManager keeps track of AIOContexts used for async IOs. AIOManager // must be cloned when CLONE_VM is used. aioManager aioManager + + // sleepForActivation indicates whether the task should report to be sleeping + // before trying to activate the address space. When set to true, delays in + // activation are not reported as stuck tasks by the watchdog. + sleepForActivation bool + + // vdsoSigReturnAddr is the address of 'vdso_sigreturn'. + vdsoSigReturnAddr uint64 } // vma represents a virtual memory area. @@ -278,7 +287,7 @@ type vma struct { mlockMode memmap.MLockMode // numaPolicy is the NUMA policy for this vma set by mbind(). - numaPolicy int32 + numaPolicy linux.NumaPolicy // numaNodemask is the NUMA nodemask for this vma set by mbind(). numaNodemask uint64 diff --git a/pkg/sentry/mm/mm_test.go b/pkg/sentry/mm/mm_test.go index edacca741..fdc308542 100644 --- a/pkg/sentry/mm/mm_test.go +++ b/pkg/sentry/mm/mm_test.go @@ -31,7 +31,7 @@ import ( func testMemoryManager(ctx context.Context) *MemoryManager { p := platform.FromContext(ctx) mfp := pgalloc.MemoryFileProviderFromContext(ctx) - mm := NewMemoryManager(p, mfp) + mm := NewMemoryManager(p, mfp, false) mm.layout = arch.MmapLayout{ MinAddr: p.MinUserAddress(), MaxAddr: p.MaxUserAddress(), diff --git a/pkg/sentry/mm/procfs.go b/pkg/sentry/mm/procfs.go index 1ab92f046..6efe5102b 100644 --- a/pkg/sentry/mm/procfs.go +++ b/pkg/sentry/mm/procfs.go @@ -148,7 +148,7 @@ func (mm *MemoryManager) appendVMAMapsEntryLocked(ctx context.Context, vseg vmaI // Do not include the guard page: fs/proc/task_mmu.c:show_map_vma() => // stack_guard_page_start(). - fmt.Fprintf(b, "%08x-%08x %s%s %08x %02x:%02x %d ", + lineLen, _ := fmt.Fprintf(b, "%08x-%08x %s%s %08x %02x:%02x %d ", vseg.Start(), vseg.End(), vma.realPerms, private, vma.off, devMajor, devMinor, ino) // Figure out our filename or hint. @@ -165,7 +165,7 @@ func (mm *MemoryManager) appendVMAMapsEntryLocked(ctx context.Context, vseg vmaI } if s != "" { // Per linux, we pad until the 74th character. - if pad := 73 - b.Len(); pad > 0 { + if pad := 73 - lineLen; pad > 0 { b.WriteString(strings.Repeat(" ", pad)) } b.WriteString(s) diff --git a/pkg/sentry/mm/syscalls.go b/pkg/sentry/mm/syscalls.go index c5dfa5972..3f496aa9f 100644 --- a/pkg/sentry/mm/syscalls.go +++ b/pkg/sentry/mm/syscalls.go @@ -974,7 +974,7 @@ func (mm *MemoryManager) MLockAll(ctx context.Context, opts MLockAllOpts) error } // NumaPolicy implements the semantics of Linux's get_mempolicy(MPOL_F_ADDR). -func (mm *MemoryManager) NumaPolicy(addr usermem.Addr) (int32, uint64, error) { +func (mm *MemoryManager) NumaPolicy(addr usermem.Addr) (linux.NumaPolicy, uint64, error) { mm.mappingMu.RLock() defer mm.mappingMu.RUnlock() vseg := mm.vmas.FindSegment(addr) @@ -986,7 +986,7 @@ func (mm *MemoryManager) NumaPolicy(addr usermem.Addr) (int32, uint64, error) { } // SetNumaPolicy implements the semantics of Linux's mbind(). -func (mm *MemoryManager) SetNumaPolicy(addr usermem.Addr, length uint64, policy int32, nodemask uint64) error { +func (mm *MemoryManager) SetNumaPolicy(addr usermem.Addr, length uint64, policy linux.NumaPolicy, nodemask uint64) error { if !addr.IsPageAligned() { return syserror.EINVAL } diff --git a/pkg/sentry/mm/vma.go b/pkg/sentry/mm/vma.go index 9a14e69e6..16d8207e9 100644 --- a/pkg/sentry/mm/vma.go +++ b/pkg/sentry/mm/vma.go @@ -195,7 +195,7 @@ func (mm *MemoryManager) applicationAddrRange() usermem.AddrRange { // Preconditions: mm.mappingMu must be locked. func (mm *MemoryManager) findLowestAvailableLocked(length, alignment uint64, bounds usermem.AddrRange) (usermem.Addr, error) { - for gap := mm.vmas.LowerBoundGap(bounds.Start); gap.Ok() && gap.Start() < bounds.End; gap = gap.NextGap() { + for gap := mm.vmas.LowerBoundGap(bounds.Start); gap.Ok() && gap.Start() < bounds.End; gap = gap.NextLargeEnoughGap(usermem.Addr(length)) { if gr := gap.availableRange().Intersect(bounds); uint64(gr.Length()) >= length { // Can we shift up to match the alignment? if offset := uint64(gr.Start) % alignment; offset != 0 { @@ -214,7 +214,7 @@ func (mm *MemoryManager) findLowestAvailableLocked(length, alignment uint64, bou // Preconditions: mm.mappingMu must be locked. func (mm *MemoryManager) findHighestAvailableLocked(length, alignment uint64, bounds usermem.AddrRange) (usermem.Addr, error) { - for gap := mm.vmas.UpperBoundGap(bounds.End); gap.Ok() && gap.End() > bounds.Start; gap = gap.PrevGap() { + for gap := mm.vmas.UpperBoundGap(bounds.End); gap.Ok() && gap.End() > bounds.Start; gap = gap.PrevLargeEnoughGap(usermem.Addr(length)) { if gr := gap.availableRange().Intersect(bounds); uint64(gr.Length()) >= length { // Can we shift down to match the alignment? start := gr.End - usermem.Addr(length) |